does Q4 quantization add noise to the workspace?

[solarkyle]

run the same trivia set on a bf16 vs 4bit quantized model with the same lens, compare smear. also tells us if the router survives quantized inference, which is how everyone actually runs local models